1 /* SPDX-License-Identifier: GPL-2.0 */ 2 /* Copyright(c) 2020 Intel Corporation. */ 3 4 #ifndef XSK_BUFF_POOL_H_ 5 #define XSK_BUFF_POOL_H_ 6 7 #include <linux/if_xdp.h> 8 #include <linux/types.h> 9 #include <linux/dma-mapping.h> 10 #include <linux/bpf.h> 11 #include <net/xdp.h> 12 13 struct xsk_buff_pool; 14 struct xdp_rxq_info; 15 struct xsk_queue; 16 struct xdp_desc; 17 struct xdp_umem; 18 struct xdp_sock; 19 struct device; 20 struct page; 21 22 #define XSK_PRIV_MAX 24 23 24 struct xdp_buff_xsk { 25 struct xdp_buff xdp; 26 u8 cb[XSK_PRIV_MAX]; 27 dma_addr_t dma; 28 dma_addr_t frame_dma; 29 struct xsk_buff_pool *pool; 30 u64 orig_addr; 31 struct list_head free_list_node; 32 struct list_head xskb_list_node; 33 }; 34 35 #define XSK_CHECK_PRIV_TYPE(t) BUILD_BUG_ON(sizeof(t) > offsetofend(struct xdp_buff_xsk, cb)) 36 37 struct xsk_dma_map { 38 dma_addr_t *dma_pages; 39 struct device *dev; 40 struct net_device *netdev; 41 refcount_t users; 42 struct list_head list; /* Protected by the RTNL_LOCK */ 43 u32 dma_pages_cnt; 44 bool dma_need_sync; 45 }; 46 47 struct xsk_buff_pool { 48 /* Members only used in the control path first. */ 49 struct device *dev; 50 struct net_device *netdev; 51 struct list_head xsk_tx_list; 52 /* Protects modifications to the xsk_tx_list */ 53 spinlock_t xsk_tx_list_lock; 54 refcount_t users; 55 struct xdp_umem *umem; 56 struct work_struct work; 57 struct list_head free_list; 58 struct list_head xskb_list; 59 u32 heads_cnt; 60 u16 queue_id; 61 62 /* Data path members as close to free_heads at the end as possible. */ 63 struct xsk_queue *fq ____cacheline_aligned_in_smp; 64 struct xsk_queue *cq; 65 /* For performance reasons, each buff pool has its own array of dma_pages 66 * even when they are identical. 67 */ 68 dma_addr_t *dma_pages; 69 struct xdp_buff_xsk *heads; 70 struct xdp_desc *tx_descs; 71 u64 chunk_mask; 72 u64 addrs_cnt; 73 u32 free_list_cnt; 74 u32 dma_pages_cnt; 75 u32 free_heads_cnt; 76 u32 headroom; 77 u32 chunk_size; 78 u32 chunk_shift; 79 u32 frame_len; 80 u8 tx_metadata_len; /* inherited from umem */ 81 u8 cached_need_wakeup; 82 bool uses_need_wakeup; 83 bool dma_need_sync; 84 bool unaligned; 85 void *addrs; 86 /* Mutual exclusion of the completion ring in the SKB mode. Two cases to protect: 87 * NAPI TX thread and sendmsg error paths in the SKB destructor callback and when 88 * sockets share a single cq when the same netdev and queue id is shared. 89 */ 90 spinlock_t cq_lock; 91 struct xdp_buff_xsk *free_heads[]; 92 }; 93 94 /* Masks for xdp_umem_page flags. 95 * The low 12-bits of the addr will be 0 since this is the page address, so we 96 * can use them for flags. 97 */ 98 #define XSK_NEXT_PG_CONTIG_SHIFT 0 99 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT) 100 101 /* AF_XDP core. */ 102 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs, 103 struct xdp_umem *umem); 104 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev, 105 u16 queue_id, u16 flags); 106 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs, 107 struct net_device *dev, u16 queue_id); 108 int xp_alloc_tx_descs(struct xsk_buff_pool *pool, struct xdp_sock *xs); 109 void xp_destroy(struct xsk_buff_pool *pool); 110 void xp_get_pool(struct xsk_buff_pool *pool); 111 bool xp_put_pool(struct xsk_buff_pool *pool); 112 void xp_clear_dev(struct xsk_buff_pool *pool); 113 void xp_add_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs); 114 void xp_del_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs); 115 116 /* AF_XDP, and XDP core. */ 117 void xp_free(struct xdp_buff_xsk *xskb); 118 119 static inline void xp_init_xskb_addr(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool, 120 u64 addr) 121 { 122 xskb->orig_addr = addr; 123 xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom; 124 } 125 126 static inline void xp_init_xskb_dma(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool, 127 dma_addr_t *dma_pages, u64 addr) 128 { 129 xskb->frame_dma = (dma_pages[addr >> PAGE_SHIFT] & ~XSK_NEXT_PG_CONTIG_MASK) + 130 (addr & ~PAGE_MASK); 131 xskb->dma = xskb->frame_dma + pool->headroom + XDP_PACKET_HEADROOM; 132 } 133 134 /* AF_XDP ZC drivers, via xdp_sock_buff.h */ 135 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq); 136 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev, 137 unsigned long attrs, struct page **pages, u32 nr_pages); 138 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs); 139 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool); 140 u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max); 141 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count); 142 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr); 143 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr); 144 static inline dma_addr_t xp_get_dma(struct xdp_buff_xsk *xskb) 145 { 146 return xskb->dma; 147 } 148 149 static inline dma_addr_t xp_get_frame_dma(struct xdp_buff_xsk *xskb) 150 { 151 return xskb->frame_dma; 152 } 153 154 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb); 155 static inline void xp_dma_sync_for_cpu(struct xdp_buff_xsk *xskb) 156 { 157 xp_dma_sync_for_cpu_slow(xskb); 158 } 159 160 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma, 161 size_t size); 162 static inline void xp_dma_sync_for_device(struct xsk_buff_pool *pool, 163 dma_addr_t dma, size_t size) 164 { 165 if (!pool->dma_need_sync) 166 return; 167 168 xp_dma_sync_for_device_slow(pool, dma, size); 169 } 170 171 /* Masks for xdp_umem_page flags. 172 * The low 12-bits of the addr will be 0 since this is the page address, so we 173 * can use them for flags. 174 */ 175 #define XSK_NEXT_PG_CONTIG_SHIFT 0 176 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT) 177 178 static inline bool xp_desc_crosses_non_contig_pg(struct xsk_buff_pool *pool, 179 u64 addr, u32 len) 180 { 181 bool cross_pg = (addr & (PAGE_SIZE - 1)) + len > PAGE_SIZE; 182 183 if (likely(!cross_pg)) 184 return false; 185 186 return pool->dma_pages && 187 !(pool->dma_pages[addr >> PAGE_SHIFT] & XSK_NEXT_PG_CONTIG_MASK); 188 } 189 190 static inline bool xp_mb_desc(struct xdp_desc *desc) 191 { 192 return desc->options & XDP_PKT_CONTD; 193 } 194 195 static inline u64 xp_aligned_extract_addr(struct xsk_buff_pool *pool, u64 addr) 196 { 197 return addr & pool->chunk_mask; 198 } 199 200 static inline u64 xp_unaligned_extract_addr(u64 addr) 201 { 202 return addr & XSK_UNALIGNED_BUF_ADDR_MASK; 203 } 204 205 static inline u64 xp_unaligned_extract_offset(u64 addr) 206 { 207 return addr >> XSK_UNALIGNED_BUF_OFFSET_SHIFT; 208 } 209 210 static inline u64 xp_unaligned_add_offset_to_addr(u64 addr) 211 { 212 return xp_unaligned_extract_addr(addr) + 213 xp_unaligned_extract_offset(addr); 214 } 215 216 static inline u32 xp_aligned_extract_idx(struct xsk_buff_pool *pool, u64 addr) 217 { 218 return xp_aligned_extract_addr(pool, addr) >> pool->chunk_shift; 219 } 220 221 static inline void xp_release(struct xdp_buff_xsk *xskb) 222 { 223 if (xskb->pool->unaligned) 224 xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb; 225 } 226 227 static inline u64 xp_get_handle(struct xdp_buff_xsk *xskb) 228 { 229 u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start; 230 231 offset += xskb->pool->headroom; 232 if (!xskb->pool->unaligned) 233 return xskb->orig_addr + offset; 234 return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT); 235 } 236 237 #endif /* XSK_BUFF_POOL_H_ */ 238